The AI review rates the findings as likely false positive, but the risk score (63/100) still counts them. The review covers version 1.21.2, not the analysed version 1.26.0.
Analysis record
- Analysed
- 2 days ago 20 November 2026
- Version
- v1.26.0
- Artifact
- SHA256 0ab4d93c…fb3539
- Detail
- Full list
Is @kya-os/mcp safe?
@kya-os/mcp describes itself as a reference implementation of the KYA-OS protocol for MCP servers, covering identity, delegation, proofs, sessions and verifiable audit trails. It declares no special permissions and no host permissions, which fits a library other servers import rather than a standalone service. The outbound calls all sit inside shipped files such as dist/authz/oidc/oidc-adapter.js and dist/delegation/did-web-resolver.js, where they fetch OIDC metadata and DID documents. That work demands network access by design, and the same pattern repeats across dist/integrations/cheqd and dist/providers/runtime-fetch.js.
The single critical match is MCP-TRANSPORT-HARDCODED-TOKEN-dist/authz/examples/inmemory-oidc.js. Read the path and the finding loses its teeth. The file sits under a directory named examples and implements an in-memory OIDC provider used to test the authorization flow, so a fixed signing key or client secret there gives the test suite something to authenticate against. A real leak would put that credential in the production authorization path, not in a demo provider that never contacts a live identity service.
Nothing in the scan turned up hidden instructions aimed at an AI agent, and the package never reads credential stores such as .ssh, .aws or .kube. We found no tool-poisoning patterns at all, which matters more than usual here because the code registers protocol primitives instead of tool descriptions.
The scanner flagged the example token by matching a literal secret string. That rule cannot tell a test fixture from a live key, and the file location settles which one this is.
Evidence ledger
Ranked by severity · findings with a source location link to the code viewer
Finding Categories
MCP Server Analysis
MCP servers expose tools and resources to AI assistants. Unlike browser extensions, they run as standalone processes with direct system access. Tool definitions are analyzed for prompt injection, data exfiltration, and tool poisoning patterns.
AI Security Report
AI Security Review
Evidence context: threat category none; evidence quality moderate.
@kya-os/mcp bills itself as a reference implementation of the KYA-OS protocol for MCP servers, covering identity, delegation, proofs, sessions and audit trails. Every finding lands inside the compiled dist tree, and the package declares no browser permissions and no host permissions.
Tool poisoning. The scan matched no tool-poisoning patterns, so there are no directives aimed at an AI agent, no invisible Unicode in tool metadata and no instruction tags. That absence carries weight here because this package defines protocol primitives (identity cards, delegation chains, session proofs) rather than tool descriptions, so there is little description surface to poison in the first place.
Credential scope. The one critical hit is MCP-TRANSPORT-HARDCODED-TOKEN-dist/authz/examples/inmemory-oidc.js. The path matters. The file lives under dist/authz/examples and holds an in-memory OIDC provider, the sort of test double that ships with a fixed signing key or client secret so integration tests can authenticate without an external identity provider. The transport rule matches a literal token string and cannot separate a fixture from a live credential. Nothing else in the package reads .ssh, .aws, .kube, application_default_credentials.json or any other credential store, and no environment harvesting loop turned up.
Network destinations. The NET-FETCH findings cluster in files whose names describe their purpose: dist/authz/oidc/oidc-adapter.js, dist/delegation/did-web-resolver.js, dist/card/resolve.js, dist/card/revocation.js, dist/integrations/cheqd/registrar.js, dist/integrations/cheqd/resolver.js and dist/providers/runtime-fetch.js. Those files resolve OIDC metadata, DID documents, credential status lists and cheqd registry entries. A protocol that verifies delegation and revocation has to reach the network to do it, and the provider abstraction in dist/providers/runtime-fetch.js exists to route those calls. No finding pairs a credential read with a POST to an unrelated domain, which is the shape exfiltration takes. The package lists no endpoints of its own because it talks to whatever issuer and resolver the operator configures.
Strongest counterargument. A hardcoded token in a shipped artifact becomes a real secret exposure if that token also grants access to something live. Someone could argue the examples directory ends up in node_modules and the credential is therefore active. That reading falls apart on two details. The file belongs to an in-memory provider whose whole job is to stand in for a real IdP during tests, and no other file in the package reads that value for authentication against a remote service. The finding is a packaging hygiene issue worth cleaning up, not evidence of theft.
Overall the evidence points to a library doing cryptographic protocol work and reaching the network only where the protocol requires it.
Key Reasons
- Zero tool-poisoning matches across the package, consistent with a library that defines protocol primitives rather than AI-facing tool descriptions
- The only critical finding, MCP-TRANSPORT-HARDCODED-TOKEN-dist/authz/examples/inmemory-oidc.js, sits in an examples directory implementing an in-memory OIDC test provider
- All NET-FETCH findings land in purpose-named files (oidc-adapter.js, did-web-resolver.js, cheqd/resolver.js, runtime-fetch.js) that resolve identity and revocation data by design
- No reads of .ssh, .aws, .kube or application_default_credentials.json anywhere in the scan
- No combination of credential reads with POSTs to unfamiliar domains
False Positive Considerations
- Literal-string secret matcher firing on a test fixture under dist/authz/examples
- NET-FETCH rules flagging expected HTTPS calls to configurable OIDC and DID resolvers
- Bundled compiled output in dist/ producing findings against transpiled library code
Reviewed 2026-10-30 · version 1.21.2; recommended action: monitor; model confidence 78%.
MCP version history
Risk trend by version
25 analyzed versions. Each point is the latest successful scan for that version; failed zero-score scans are hidden. Dates: First seen by Risky Plugins.
Pick any point on the chart, then open the Source tab to read that version's code.
Source Code Not Available
Source code is not available for this version of the extension.
About This Extension
Frequently Asked Questions
Similar Extensions
Related extensions from the same publisher or marketplace